The useful lesson is more specific than the headline claim that AI doubled administrative time savings: employees can bring orders from different communication channels into a conversational workflow, consult existing business records, and approve the resulting transaction before it is created.
What the time-savings figures actually mean
Microsoft’s account contains two related but distinct measurements:
- Business Central’s existing administrative savings: Owner Elisabeth Colman estimates the platform already saved the company between one and 1.5 hours daily.
- Order-entry work after adding agents: Colman estimates that a task previously occupying one to two hours each day now takes less than 30 minutes.
Microsoft summarizes the combined outcome as more than doubling the time savings previously achieved with Business Central. However, the detailed figures measure different scopes of work; they should not be added together to manufacture a precise daily productivity result. These are customer-reported estimates in a Microsoft case study, not independently audited measurements.
That distinction does not erase the operational improvement. It simply keeps a promising example from becoming a universal claim that installing an agent doubles everyone’s productivity.
Why a caviar business needed custom agents
Imperial Heritage uses Business Central for finance, sales, customer management, purchasing, inventory, and supply-chain activities. It manages approximately 500 active customers, with orders arriving through WhatsApp, email, SMS, and telephone calls. Some chefs place orders after evening service, leaving staff with requests to process before deliveries can be prepared.
The bottleneck was the handoff between those conversations and the structured records needed to fulfill an order.
Imperial Heritage and implementation partner Deltanexx considered Business Central’s prebuilt Sales Order Agent, but wanted a workflow accommodating multiple channels and employees’ preferred AI assistants. Deltanexx therefore developed two custom agents: one using Copilot Studio and another using Claude. Both connect to Business Central through its Model Context Protocol, or MCP, server. Microsoft says the Copilot Studio implementation also provides access through Teams and Microsoft 365 Copilot.
There is a meaningful distinction here. Microsoft Learn describes the built-in Sales Order Agent as an email-centered workflow: it identifies a customer contact, prepares a quote for review and sending, and converts the quote into an order after the customer confirms it. Imperial Heritage’s custom implementation addresses a different intake pattern rather than demonstrating that every business needs to replace the built-in agent.
From a customer message to an approved order
The implementation gives each agent three core skills: creating sales orders, creating customer records, and preparing customer follow-up reports. The sales-order workflow described by Microsoft works as follows:
- An employee supplies the request. They can paste a WhatsApp message into the conversation or describe an order received by SMS or telephone using voice.
- The agent identifies the relevant records. It matches the request to customers and items in Business Central.
- Existing transactions provide context. An employee can ask it to consult a previous invoice for agreed pricing, product codes, and other details.
- The employee reviews the proposal. The agent presents the order details and requests confirmation before creating the order.
This is assisted multichannel order entry, not evidence of an unattended system automatically ingesting every message and call. The described examples still involve employees bringing information into the agent conversation.
For customer creation, the agent checks whether a record already exists before creating another. Reporting capabilities include customer revenue, outstanding orders, weekly follow-up reports, and daily or weekly business snapshots. Microsoft reports that the custom solution took a few weeks to develop.
The practical design choice is worth noting: conversational input sits in front of structured ERP transactions, with a review step between interpretation and execution. The chatbot is not being handed the company keys and wished good luck.
The administrator’s question: what can it change?
Microsoft describes several controls in this deployment. Copilot Studio uses the Business Central MCP tool and Microsoft’s pre-registered application. The Claude connection uses a separate application registered in Imperial Heritage’s Microsoft Entra ID tenant, with administrator consent for delegated Business Central access. MCP configuration defines available APIs and operations, while Azure Application Insights records calls for monitoring and support.
Microsoft’s separate MCP configuration documentation adds an important deployment boundary: the documented server applies to Business Central online. It also distinguishes read access from permission to create, modify, delete, or invoke actions. Those write operations must be configured; connecting an assistant does not, by itself, authorize every possible transaction.
For administrators evaluating a similar project, the practical implication is to assess the permitted operations as carefully as the conversational experience. Reading previous invoices, creating orders, deleting records, and posting documents are different capabilities—not one interchangeable “AI access” permission.
The case study does not establish data-retention periods, third-party processing terms, or a complete security assessment. Its reported controls are useful implementation details, not a blanket assurance about every MCP-connected assistant.
What is live—and what remains a plan
Microsoft says all incoming orders now receive agent support. Colman also reports fewer mistakes, particularly with complex orders and special pricing, but the story supplies no measured error rate or independent accuracy assessment.
Further purchasing and analytical capabilities remain prospective. Vendor and purchase-order APIs are already included in the MCP configuration, but Imperial Heritage is exploring additional purchasing skills, identifying regular customers who have not reordered, and assistance with supplier price increases and margins. Available APIs should not be confused with completed business workflows.
The strongest takeaway is therefore not that AI has made administration disappear. It is that a narrowly defined, human-reviewed workflow can reduce the effort of moving everyday customer conversations into Business Central. For other businesses considering agents, that is the more useful starting point: identify the repetitive handoff, constrain what the assistant can do, and measure the work it actually changes.